Shadowing Practice: How might LLMs store facts | Deep Learning Chapter 7 - Learn English Speaking with Video

If you feed a large language model the phrase, Michael Jordan plays the sport of blank, and you have it predict what comes next, and it correctly predicts basketball,

this would suggest that somewhere, inside its hundreds of billions of parameters, it's baked in knowledge about a specific person and his specific sport.

And I think in general, anyone who's played around with one of these models has the clear sense that it's memorized tons and tons of facts.

So a reasonable question you could ask is, how exactly does that work? work, and where do those facts live?

Last December, a few researchers from Google DeepMind posted about work on this question, and they were using this specific example of matching athletes to their sports.

And although a full mechanistic understanding of how facts are stored remains unsolved, they had some interesting partial results,

including the very general high -level conclusion that the facts seem to live inside a specific part of these networks,

known fancifully as the Multilayer Perceptrons, or MLPs for short.

In the last couple of chapters, you and I have been digging into the details behind Transformers, the architecture underlying large language models, and also underlying a lot of other modern AI.

In the most recent chapter, we were focusing on a piece called Attention.

And the next step, for you and me, is to dig into the details of what happens inside these multi -layer perceptrons, which make up the other big portion of the network.

The computation here is actually relatively simple, especially when you compare it to a tension.

It boils down essentially to a pair of matrix multiplications with a simple something in between.

However, interpreting what these computations are doing is exceedingly challenging.

Our main goal here is to step through the computations and make them memorable, but I'd like to do it in the context of showing a specific example of how one of these blocks could,

at least in principle, store a concrete fact.

Specifically, it'll be storing the fact that Michael Jordan plays basketball.

I should mention the layout here is inspired by a conversation I had with one of those DeepMind researchers, Neil Nanda.

For the most part, I will assume that you've either watched the last two chapters, or otherwise you have a basic sense for what a transformer is.

But refreshers never hurt, so here's the quick reminder of the overall flow.

You and I have been studying a model that's trained to take in a piece of text and predict what comes next.

That input text is first broken into a bunch of tokens, which means little chunks that are typically words or little pieces of words,

and each token is associated with a high -dimensional vector, which is to say a long list of numbers.

This sequence of vectors then repeatedly passes through two kinds of operation.

Attention, which allows the vectors to pass information between one another, and then the multi -layer perceptrons, the thing that we're to dig into today.

And also there's a certain normalization step in between.

After the sequence of vectors has flowed through many many different iterations of both of these blocks,

by the end, the hope is that each vector has soaked up enough information, both from the context ,

and also from the general knowledge that was baked into the model weights through training,

that it can be used to make a prediction of what token comes next.

One of the key ideas that I want you to have in your mind is

that all of these vectors live in a very very high dimensional space, and when you think about that space, different directions can encode different kinds of meaning.

So a very classic example that I like to refer back to is how

if you look at the embedding of woman and subtract the embedding of man, and you take that little step and you add it to another masculine noun,

something like uncle, you land somewhere very very close to the corresponding feminine noun.

In this sense, this particular direction encodes gender information.

The idea is

that many other distinct directions in this super high -dimensional space could correspond to other features

that the model might want to represent.

In a transformer, these vectors don't merely encode the meaning of a single word, though.

As they flow through the network, they imbibe a much richer meaning based on all the context around them, and also based on the model's knowledge.

Ultimately, each one needs to encode something far far beyond the meaning of a single word, since it needs to be sufficient to predict what will come next.

We've already seen how attention blocks let you incorporate context, but a majority of the model parameters actually live inside the MLP blocks,

and one thought for what they might be doing is that they offer extra capacity to store facts.

Like I said, the lesson here is going to center on the concrete toy example of how exactly it could store the fact

that Michael Jordan plays basketball.

Now this toy example is going to require that you and I make a couple of assumptions about that high dimensional space.

First, we'll suppose that one of the directions represents the idea of a first name Michael.

And then another, nearly perpendicular direction, represents the idea of the last name Jordan.

And then yet a third direction will represent the idea of basketball.

So specifically what I mean by this is

if you look in the network and you pluck out one of the vectors being processed, if its dot product with this first name Michael direction is 1,

that's what it would mean for the vector to be encoding the idea of a person with that first name.

Otherwise, that dot product would be zero or negative, meaning the vector doesn't really align with that direction.

And for simplicity, let's completely ignore the very reasonable question of what it might mean if that dot product was bigger than one.

Similarly, its dot product with these other directions would tell you whether it represents the last name Jordan, or basketball.

So let's say a vector is meant to represent the full name, Michael Jordan, then its dot product with both of these directions would have to be one.

Since the text, Michael Jordan, spans two different tokens, this would also mean we have to assume

that an earlier attention block has successfully passed information to the second of these two vectors

so as to ensure that it can encode both names.

With all of those as the assumptions, let's now dive into the meat of the lesson.

What happens inside a multilayer perceptron?

You might think of this sequence of vectors flowing into the block.

And remember, each vector was originally associated with one of the tokens from the input text.

What's going to happen is that each individual vector from that sequence goes through a short series of operations, we'll unpack them in just a moment,

and at the end we'll get another vector with the same dimension.

That other vector is going to get added to the original one that flowed in, and that sum is the result flowing out.

This sequence of operations is something you apply to every vector

in the sequence associated with every token in the input and it all happens in parallel.

In particular, the vectors don't talk to each other in this step.

They're all kind of doing their own thing.

And for you and me, that actually makes it a lot simpler because it means

if we understand what happens to just one of the vectors through this block, we effectively understand what happens to all of them.

When I say this block is going to encode the fact that Michael Jordan plays basketball, what I mean is that if a vector flows in that encodes first name Michael and last name Jordan,

then this sequence of computations will produce something that includes that direction basketball, which is what will add on to the vector in that position.

The first step of this process looks like multiplying that vector by a very big matrix, no surprises there, this is deep learning.

And this matrix, like all of the other ones we've seen, is filled with model parameters that are learned from data, which you might think of as a bunch of knobs

and dials that get tweaked and tuned to determine what the model behavior is.

One nice way to think about matrix multiplication is to imagine each row of that matrix as being its own vector,

and taking a bunch of dot products between those rows and the vector being processed, which I'll label as E for embedding.

For example, suppose that very first row happened to equal this first name Michael direction that we're presuming exists.

That would mean that the first component in this output, this dot product right here, would be 1 if that vector encodes the first name Michael,

and 0 or negative otherwise.

Even more fun, take a moment to think about what it would mean

if that first row was this first name Michael plus last name Jordan direction.

And for simplicity, let me go ahead and write that down as m plus j.

Then taking a dot product with this embedding e, things distribute really nicely, so it looks like m dot e plus j dot e,

and notice how that means the ultimate value would be 2 if the vector encodes the full name, Michael Jordan, and otherwise it would be 1 or something smaller than 1.

And that's just one row in this matrix.

You might think of all of the other rows as, in parallel, asking some other kinds of questions, probing at some other sorts of features of the vector being processed.

Very often this step also involves adding another vector to the output, which is full of model parameters learned from data.

This other vector is known as the bias.

For our example, I want you to imagine that the value of this bias in that very first component is negative 1,

meaning our final output looks like that relevant dot product, but minus 1.

You might very reasonably ask why I would want you to assume that the model has learned this, and in a moment you'll see why it's very clean

and nice if we have a value here which is positive

if and only if our vector encodes the full name Michael Jordan, and otherwise it's 0 or negative.

The total number of rows in this matrix, which is something like the number of questions being asked, in the case of GPT -3, whose numbers we've been following,

is just under 50 ,000.

100

In fact, it's exactly four times the number of dimensions in this embedding space.

101

That's a design choice, you could make it more, you could make it less, but having a clean multiple tends to be friendly for hardware.

102

Since this matrix full of weights maps us into a higher dimensional space, I'm going to give it the shorthand.. w up.

103

I'll continue labeling the vector we're processing as e, and let's label this bias vector as b up and put that all back down in the diagram.

104

At this point, a problem is that this operation is purely linear, but language is a very non -linear process.

105

If the entry that we're measuring is high for Michael plus Jordan, it would also necessarily be somewhat triggered by Michael plus Phelps, and also Alexis plus Jordan,

106

despite those being unrelated conceptually.

107

What you really want is a simple yes or no for the full name.

108

So the next step is to pass this large intermediate vector through a very simple non -linear function.

109

A common choice is one that takes all of the negative values and maps them to zero, and leaves all of the positive values unchanged.

110

And continuing with the deep learning tradition of overly fancy names, this very simple function is often called the rectified linear unit, or or ReLU for short,

111

here's what the graph looks like.

112

So taking our imagined example where this first entry of the intermediate vector is 1

113

if and only if the full name is Michael Jordan and 0 or negative otherwise, after you pass it through the ReLU,

114

you end up with a very clean value where all of the 0 and negative values just get clipped to 0.

115

So this output would be 1 for the full name Michael Jordan and 0 otherwise.

116

In other words, it very directly mimics the behavior of an AND gate.

117

Often models will use a slightly modified function that's called the JLU, which has the same basic shape, it's just a bit smoother.

118

But for our purposes, it's a little bit cleaner if we only think about the ReLU.

119

Also, when you hear people refer to the neurons of a transformer, they're talking about these values right here.

120

Whenever you see that common neural network picture with a layer of dots

121

and a bunch of lines connecting to the previous layer, which we had earlier in this series, that's typically meant to convey this combination of a linear step,

122

a matrix multiplication, followed by some simple term -wise non -linear function like a ReLU.

123

You would say that this neuron is active whenever this value is positive, and that it's inactive if that value is zero.

124

The next step looks very similar to the first one.

125

You multiply by a very large matrix, and you add on a certain bias term.

126

In this case, the number of dimensions in the output is back down to the size of that embedding space, so I'm going to go ahead and call this the down projection matrix.

127

And this time, instead of thinking of things row by row, it's actually nicer to think of it column by column.

128

You see, another way

129

that you can hold matrix multiplication in your head is to imagine taking each column of the matrix

130

and multiplying it by the corresponding term in the vector that it's processing and adding together all of those rescaled columns.

131

The reason it's nicer to think about this way is because here the columns have the same dimension as the embedding space, so we can think of them as directions in that space.

132

For instance, we will imagine that the model has learned to make

133

that first column into this basketball direction that we suppose exists.

134

What that would mean is that when the relevant neuron in that first position is active, we'll be adding this column to the final result.

135

But if that neuron was inactive, if that number was zero, then this would have no effect.

136

And it doesn't just have to be basketball.

137

The model could also bake into this column many other features

138

that it wants to associate with something that has the full name Michael Jordan.

139

And at the same time, all of the other columns in this matrix are telling you what will be added to the final result

140

if the corresponding neuron is active.

141

And if you have a bias in this case, it's something that you're just adding every single time, regardless of the neuron values.

142

You might wonder what's that doing.

143

As with all parameter -filled objects here, it's kind of hard to say exactly.

144

Maybe there's some bookkeeping that the network needs to do, but you can feel free to ignore it for now.

145

Making our notation a little more compact again, I'll call this big matrix W down, and similarly call that bias vector B down, and put that back into our diagram.

146

Like I previewed earlier, what you do with this final result is add it to the vector

147

that flowed into the block at that position, and that gets you this final result.

148

So for example, if the vector flowing in encoded both first name Michael and last name Jordan,

149

then because this sequence of operations will trigger that AND gate, it will add ON the basketball direction so what pops out will encode all of those together.

150

And remember, this is a process happening to every one of those vectors in parallel.

151

In particular, taking the GPT -3 numbers, it means that this block doesn't just have 50 ,000 neurons in it,

152

it has 50 ,000 times the number of tokens in the input.

153

So that is the entire operation, two matrix products each with a bias added, and a simple clipping function in between.

154

Any of you who watched the earlier videos of the series

155

will recognize this structure as the most basic kind of neural network that we studied there.

156

In that example, it was trained to recognize handwritten digits.

157

Over here, in the context of a transformer for a large language model, this is one piece in a larger architecture,

158

and any attempt to interpret what exactly it's doing is heavily

159

intertwined with the idea of encoding information into vectors of a high -dimensional embedding space.

160

That is the core lesson, but I do want to step back and reflect on two different things.

161

The first of which is a kind of bookkeeping, and the second of which involves a very thought provoking fact about higher dimensions

162

that I actually didn't know until I dug into Transformers.

163

In the last two chapters, you and I started counting up the total number of parameters in GPT -3 and seeing exactly where they live, so let's quickly finish up the game here.

164

I already mentioned how this up projection matrix has just under 50 ,000 rows, and that each row matches the size of the embedding space,

165

which for GPT -3 is 12 ,288.

166

Multiplying those together, it gives us 604 million parameters just for that matrix.

167

And the down projection has the same number of parameters, just with a transposed shape.

168

So together, they give about 1 .2 billion parameters.

169

The bias vector also accounts for a couple more parameters, but it's a trivial proportion of the total, so I'm not even going to show it.

170

In GPT -3, this sequence of embedding vectors flows through not one,

171

but 96 distinct MLPs, so the total number of parameters devoted to all of these blocks adds up to about 116 billion.

172

This is around two -thirds of the total parameters in the network, and when you add it to everything that we had before for the attention blocks, the embedding, and the unembedding,

173

you do indeed get that grand total of 175 billion as advertised.

174

It's probably worth mentioning there's another set of parameters associated with those normalization steps that this explanation has skipped over, but like the bias vector,

175

they account for a very trivial proportion of the total.

176

As to that second point of reflection, you might be wondering if this central toy example we've been spending

177

so much time on reflects how facts are actually stored in real large language models.

178

It is true that the rows of that first matrix can be thought of as directions in this embedding space,

179

and that means the activation of each neuron tells you how much a given vector aligns with some specific direction.

180

It's also true that the columns of

181

that second matrix tell you what will be added to the result if that neuron is active.

182

Both of those are just mathematical facts.

183

However, the evidence does suggest that individual neurons very rarely represent a single clean feature like Michael Jordan.

184

And there may actually be a very good reason this is the case, related to an idea floating around interpretability researchers these days known as superposition.

185

This is a hypothesis

186

that might help to explain both why the models are especially hard to interpret and also why they scale surprisingly well.

187

The basic idea is that if you have an n -dimensional space

188

and you want to represent a bunch of different features using directions that are all perpendicular to one another in that space, you know, that way if you add a component in one direction,

189

it doesn't influence any of the other directions, then the maximum number of vectors you can fit is only n, the number of dimensions.

190

To a mathematician, actually, this is the definition of dimension.

191

But where it gets interesting is if you relax that constraint a little bit and you tolerate some noise.

192

Say you allow those features to be represented by vectors that aren't exactly perpendicular, they're just nearly perpendicular, maybe between 89 and 91 degrees apart.

193

If we were in two or three dimensions, this makes no difference.

194

That gives you hardly any extra wiggle room to fit more vectors in, which makes it all the more counterintuitive that for higher dimensions, the answer changes dramatically.

195

I can give you a really quick

196

and dirty illustration of this using some scrappy python that's going to create a list of 100 dimensional vectors,

197

each one initialized randomly, and this list is going to contain 10 ,000 distinct vectors, so 100 times as many vectors as there are dimensions.

198

This plot right here shows the distribution of angles between pairs of these vectors.

199

So because they started at random, those angles could be anything from 0 to 180 degrees, but you'll notice that already, even just for random vectors,

200

there's this heavy bias for things to be closer to 90 degrees.

201

Then what I'm going to do is run a certain optimization process

202

that iteratively nudges all of these vectors so that they try to become more perpendicular to one another.

203

After repeating this many different times, here's what the distribution of angles looks like.

204

We have to actually zoom in on it here

205

because all of the possible angles between pairs of vectors sit inside this narrow range between 89 and 91 degrees.

206

In general, a consequence of something known as the Johnson -Lindenstrauss Lemma is

207

that the number of vectors you can cram into a space

208

that are nearly perpendicular like this grows exponentially with the number of dimensions.

209

This is very significant for large language models, which might benefit from associating independent ideas with nearly perpendicular directions.

210

It means that it's possible for it to store many, many more ideas than there are dimensions in the space that it's allotted.

211

This might partially explain why model performance seems to scale so well with size.

212

A space that has 10 times as many dimensions can store way, way more than 10 times as many independent ideas.

213

And this is relevant not just to that embedding space where the vectors flowing through the model live,

214

but also to that vector full of neurons in the middle of that multilayer perceptron that we just studied.

215

That is to say, at the sizes of GPT -3, it might not just be probing at 50 ,000 features,

216

but if it instead leveraged this enormous added capacity by using nearly perpendicular directions of the space, it could be probing at many,

217

many more features of the vector being processed.

218

But if it was doing that, what it means is that individual features aren't going to be visible as a single neuron lighting up.

219

It would have to look like some specific combination of neurons instead, a superposition.

220

For any of you curious to learn more, a key relevant search term here is sparse autoencoder, which is a tool

221

that some of the interpretability people use to try to extract what the true features are even

222

if they're very superimposed on all these neurons.

223

I'll link to a couple really great anthropic posts all about this.

224

At this point, we haven't touched every detail of a transformer, but you and I have hit the most important points.

225

The main thing that I want to cover in a next chapter is the training process.

226

On the one hand, the short answer for how training works is that it's all backpropagation, and we covered backpropagation in a separate context with earlier chapters in the series.

227

But there is more to discuss, like the specific cost function used for language models, the idea of fine -tuning using reinforcement learning with human feedback,

228

and the notion of scaling laws.

229

Quick note for the active followers among you, there are a number of non -machine learning related videos

230

that I'm excited to sink my teeth into before I make

231

that next chapter so it might be a while but I do promise it'll come in due time.

📺 Mismo canal

Essence of linear algebra preview

en-USUniversity Students+2

Linear combinations, span, and basis vectors | Chapter 2, Essence of linear algebra

en-USUniversity Students+2

✨ Recomendar Video

172 oraciones

How Is a Memory Stored Inside Your Squishy Brain?

IELTS PrepTOEFL Prep+2

105 oraciones

Google’s AI Course for Beginners (in 10 minutes)!

en-USUniversity Students+2

115 oraciones

The Unbelievable Science of How We Read

en-USIELTS Prep+3

85 oraciones

We Can Now Simulate a Human Brain, Scientists Show

IELTS PrepTOEFL Prep+2

72 oraciones

Fight brain-rot with Worked Examples

University StudentsBusiness English+1

321 oraciones

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

University StudentsWorking Adults

83 oraciones

UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?

IELTS PrepTOEFL Prep+2

7 oraciones

How AI works: Machine learning

IELTS PrepTOEFL Prep+2

146 oraciones

The Most Complex Language in the World

IELTS PrepTOEFL Prep+2

148 oraciones

[Học tiếng Anh THẬT] Elon Musk - Future, AI

IELTS PrepTOEFL Prep+2

¿Qué es la Técnica de Shadowing?

Shadowing es una técnica de aprendizaje de idiomas respaldada por la ciencia, desarrollada originalmente para la formación de intérpretes profesionales y popularizada por el políglota Dr. Alexander Arguelles. El método es simple pero poderoso: escuchas audio en inglés nativo y lo repites en voz alta de inmediato, como una sombra que sigue al hablante con solo 1-2 segundos de retraso. A diferencia de la escucha pasiva o los ejercicios de gramática, el shadowing obliga a tu cerebro y músculos de la boca a procesar y reproducir simultáneamente patrones de habla reales. Las investigaciones muestran que mejora significativamente la precisión de la pronunciación, la entonación, el ritmo, el habla conectada, la comprensión auditiva y la fluidez al hablar, convirtiéndola en una de las metodologías más efectivas para la preparación del IELTS Speaking y la comunicación en inglés en el mundo real.

Técnica de shadowing: lee la guía completa paso a paso →

Shadowing Practice: How might LLMs store facts | Deep Learning Chapter 7 - Learn English Speaking with Video

Fourier series

Essence of linear algebra preview

The inscribed square problem

Euler's Formula Poem

Linear combinations, span, and basis vectors | Chapter 2, Essence of linear algebra

The Newton art puzzle

Transformers, the tech behind LLMs | Deep Learning Chapter 5

Large Language Models explained briefly

Hilbert's Curve: Is infinite math useful?

Vectors | Chapter 1, Essence of linear algebra

But what is a neural network? | Deep learning chapter 1

Oh, wait, actually the best Wordle opener is not “crane”…

Divergence and curl: The language of Maxwell's equations, fluid flow, and more

But what is quantum computing? (Grover's Algorithm)

Tie random ends: How many loops?

How Is a Memory Stored Inside Your Squishy Brain?

Google’s AI Course for Beginners (in 10 minutes)!

The Unbelievable Science of How We Read

We Can Now Simulate a Human Brain, Scientists Show

Fight brain-rot with Worked Examples

Mixture of Experts Explained Visually: How Trillion-Parameter Models Actually Work

UniG2U-Bench: Do Unified Models Advance Multimodal Understanding?

How AI works: Machine learning

The Most Complex Language in the World

[Học tiếng Anh THẬT] Elon Musk - Future, AI

Learn English with Videos: Dive into LLMs and Boost Your Skills

Top 5 Phrases for Daily Communication (and Beyond!)

Step-by-Step Shadowing Guide for This Video

¿Qué es la Técnica de Shadowing?