Tuesday, December 21, 2021

Complex systems and AI

Abstract from the article linked in the first tweet above, "Collective Intelligence for Deep Learning: A Survey of Recent Developments":

In the past decade, we have witnessed the rise of deep learning to dominate the field of artificial intelligence. Advances in artificial neural networks alongside corresponding advances in hardware accelerators with large memory capacity, together with the availability of large datasets enabled researchers and practitioners alike to train and deploy sophisticated neural network models that achieve state-of-the-art performance on tasks across several fields spanning computer vision, natural language processing, and reinforcement learning. However, as these neural networks become bigger, more complex, and more widely used, fundamental problems with current deep learning models become more apparent. State-of-the-art deep learning models are known to suffer from issues that range from poor robustness, inability to adapt to novel task settings, to requiring rigid and inflexible configuration assumptions. Ideas from collective intelligence, in particular concepts from complex systems such as self-organization, emergent behavior, swarm optimization, and cellular systems tend to produce solutions that are robust, adaptable, and have less rigid assumptions about the environment configuration. It is therefore natural to see these ideas incorporated into newer deep learning methods. In this review, we will provide a historical context of neural network research's involvement with complex systems, and highlight several active areas in modern deep learning research that incorporate the principles of collective intelligence to advance its current capabilities. To facilitate a bi-directional flow of ideas, we also discuss work that utilize modern deep learning models to help advance complex systems research. We hope this review can serve as a bridge between complex systems and deep learning communities to facilitate the cross pollination of ideas and foster new collaborations across disciplines.

Saturday, December 18, 2021

Looking up

Substrate Independence

The notion of substrate independence in is invoked in arguments about whether or not, at least in theory, a digital computer can do anything the human brain can do. The idea is that what matters is the computational procedure, the algorithm if you will, not the substrate in which it is implemented. Anything that’s being ‘computed’ in neural ‘wetware’ can be computed by suitable software running on digital hardware. We just have to figure out the procedure and have enough computing power to do it.

I wonder, though. To the extent that the notion of substrate independence assumes (something like) the distinction between hardware and software, this may not be the case. For the hardware/software distinction doesn’t apply to the brain (and its mind). Thus, you can easily erase some software and data from a digital computer without affecting the underlying hardware. You can just as easily reload that data and software to the hardware, thus restoring the computer to its prior state. You can’t do that with a brain. Similarly, you can add new capability to computer simply by uploading new software. You can’t do that with a brain. You cannot, for example, learn a new language or a new intellectual discipline simply by uploaded a new module, say, overnight. You have to learn, painstakingly learn. In neural ‘wetware’ the substrate in irreversibly changed in a way this is not true of digital hardware.

Does machine learning change this? I note that machine learning is software that’s implemented on hardware. The underlying hardware is not changed. The learning takes place entirely in the data (the parameter weights) that is learned. Still, it seems that something more or less like (organic) neural learning is taking place in the implemented system.

Food for thought.

Neuromorphic chips and spiking neurons

From the linked article, "Spiking Neural Networks":

Artificial intelligence researchers, on the other hand, would like to build deep neural networks that have both the brain’s remarkable abilities and its extraordinary energy efficiency. The brain consumes only about 20 watts of power. If the brain achieves its ends partly because of spiking neurons, some think that energy-efficient deep artificial neural networks (ANNs) would also need to follow suit.

But spiking neural networks have been hamstrung. The very thing that made them attractive — communicating via spikes — also made them extremely difficult to train. The algorithms that ran on IBM’s chip, for instance, had to be trained offline at considerable computational cost.

That’s set to change. Researchers have developed new algorithms to train spiking neural networks that overcome some of the earlier limitations. And at least for networks of tens of thousands of neurons, these SNNs perform as well as regular ANNs. Such networks would likely be better at processing data that has a temporal dimension (such as speech and videos) compared with standard ANNs. Also, when implemented on neuromorphic chips, SNNs promise to open up a new era of low-energy, always-on devices that can operate without access to services in the cloud.

One of the key differences between a standard ANN and a spiking neural network is the model of the neuron itself. Any artificial neuron is a simplified computational model of biological neurons. A biological neuron receives inputs into the cell body via its dendrites; and based on some internal computation, the neuron may generate an output in the form of a spike on its axon, which then serves as an input to other neurons. Standard ANNs use a model of the neuron in which the information is encoded in the firing rate of the neuron. So the function that transforms the inputs into an output is often a continuous valued function that represents the spiking rate. This is achieved by first taking a weighted sum of all the inputs and then passing the sum through an activation function. For example, a sigmoid turns the weighted sum into a real value between 0 and 1.

In a spiking artificial neuron, on the other hand, the information is encoded in both the timing of the output spike and the spiking rate. The most commonly used model of such a spiking neuron in artificial neural networks is called the leaky integrate-and-fire (LIF) neuron. Input spikes cause the neuron’s membrane potential — the electrical charge across the neuron’s cell wall — to build up. There are also processes that cause this charge to leak; in the absence of input spikes, any built-up membrane potential starts to decay. But if enough input spikes come within a certain time window, then the membrane potential crosses a threshold, at which point the neuron fires an output spike. The membrane potential resets to its base value. Variations on this theme of an LIF neuron form the basic computational units of spiking neural networks.

In 1997, Wolfgang Maass of the Institute of Theoretical Computer Science, Technische Universität, Graz, Austria, showed that such SNNs are computationally more powerful, in terms of the number of neurons needed for some task, than ANNs with rate-coding neurons that use a sigmoid activation function. He also showed that SNNs and ANNs are equivalent in their ability to compute some function (an important equivalence, since an ANN’s claim to fame is that it is a universal function approximator: given some input, an ANN can be trained to approximate any function to transform the input into a desired output).

There's much more at the link, much of it having to do with how to modify backpropagation so it works and can scale with SNNs.

Conclusion:

All this bodes well for the day when spiking neural networks can be implemented on the numerous neuromorphic chips that are in development. The hope is that such networks can be both trained and deployed using dedicated hardware that sips rather than sucks energy.

Vision isn't "solved" [AI]

Abstract of linked article, Overinterpretation reveals image classification model pathologies:

Image classifiers are typically scored on their test set accuracy, but high accuracy can mask a subtle type of model failure. We find that high scoring convolutional neural networks (CNNs) on popular benchmarks exhibit troubling pathologies that allow them to display high accuracy even in the absence of semantically salient features. When a model provides a high-confidence decision without salient supporting input features, we say the classifier has overinterpreted its input, finding too much class-evidence in patterns that appear nonsensical to humans. Here, we demonstrate that neural networks trained on CIFAR-10 and ImageNet suffer from overinterpretation, and we find models on CIFAR-10 make confident predictions even when 95% of input images are masked and humans cannot discern salient features in the remaining pixel-subsets. We introduce Batched Gradient SIS, a new method for discovering sufficient input subsets for complex datasets, and use this method to show the sufficiency of border pixels in ImageNet for training and testing. Although these patterns portend potential model fragility in real-world deployment, they are in fact valid statistical patterns of the benchmark that alone suffice to attain high test accuracy. Unlike adversarial examples, overinterpretation relies upon unmodified image pixels. We find ensembling and input dropout can each help mitigate overinterpretation.

Tuesday, December 7, 2021

Thursday, December 2, 2021

It Shook Me, the Light

Bump! This post is from 2010, but it's been on my mind lately. So I'm bumping it to the head of the queue.
 
During the early 1970s I'd played for two years with a rock band called St. Matthew Passion. Modeled on Blood, Sweat, and Tears and on Chicago, the band consisted of 4-piece rhythm section plus three horns: sax, trumpet (me), and trombone. On “She's Not There” the three horns would start with a chaotic improvised freak-out and then, on cue from the keyboard player, the entire band would come in on the first bar of the written arrangement.

On our last gig it was just me and the sax player; the trombonist couldn't make it. The sax and I started our improv. The music got more and more intense until Wham! I felt myself dissolve into white light and pure music. It felt good.

And I tensed up.

It was over.

After the gig the sax player and I made a few remarks about it — “that was nice” — enough to confirm that something had happened to him too. One guy from the audience came up to us and remarked on how fine that section had been. Did he know what had happened? Or, if not ‘know’ exactly, did he sense a special magic in the performance? I ask because performers and audience often have a very different ‘sense’ of the same performance. Perhaps the guy was just complimenting us on our ‘freak-out’ chops, not on any magic in the music.

That's the only time I've ever experienced that kind of ego loss in music. For a few years I was very ambivalent about that experience, wanting it again, but also fearing it. A child of the 60s, a very geekish child of the 60s, I’d read quite a bit about altered states of consciousness, as they were called in the scientific literature. I read around in the secondary and tertiary literature on mystical experiences, and even a bit of the primary literature – though just exactly what’s the point of reading a mystic’s account an ineffable encounter with . . . . well, with what, exactly?

I knew such things happened. And now, in little more than a couple of heartbeats, now I too knew. But what is it that I knew?

Other than the experience itself, I knew that what all those people had been writing about was real. It’s not that I doubted it. Still it’s one thing to read about walking on the moon, even to see video footage and photographs of space-suited men walking about. It’s another thing to be there oneself.

But how can one experience be so powerful, so polarizing, that it haunts your thoughts and echoes through your soul for years afterward? What is the human nervous system that THAT can happen? In attributing the experience to the nervous system – as opposed, say, to an encounter with the divine, I do not thereby mean to dismiss it – oh, that? that was just a burp of the nervous system. We cannot dismiss it. The nervous system is us.

Now the memory's faded & the ambivalence too. But I'm playing better music now than ever I did back then. I’m talking not so much about technique – that comes and goes – but about expressive power, about ‘authenticity.’ Is that authenticity and echo of that experience?

Who knows?


* * * * *


Such experiences are common enough among musicians. Over the years I’ve collected accounts from books and articles. Here are some from Jenny Boyd, with Holly George-Warren. Musicians In Tune. New York: Simon and Schuster, 1992.
 
Patty Smyth (p. 161):
When I have had those experiences, I'm singing by myself. I've had those moments when I do feel the voice coming through me, and I know it's coming from out there. It's a certain tone in that voice that makes me feel that way. It chokes me up.
Cece Bullard (pp. 161-162):
It's like you leave your body. It's like you're dizzy and lightheaded and yet right there. My hands just seem to throb, like a pulse almost. It's the best feeling in the world, bar none. It took me a lot of singing lessons before I finally connected with that feeling. The first time it clicked and I connected, I nearly fell down, and I started crying.
Sinéad O'Conner (p. 164):
A lot of times I shake uncontrollably. I can't control the shaking, and it's not because I'm nervous, it's because I'm singing. It's because it's coming out and it's making me shake. It feels like being drunk, it's like an out-of-body experience. There are times when I've done gigs — and it doesn't happen every time you do a show or every time you write something — but they've told me stuff I've done onstage that I'm not aware I've done.
Branford Marsalis (p. 173):
High, you feel high. It's easy to do it physically, but it's hard to do it mentally. I feel that musicians who say it happens every time they play are full of shit. The sublime cannot be routine. Three times, and you never forget them. It's with a combination of musicians, it's never just me.
Ringo Starr (p. 176):
It feels great; its just a knowing. It's magic actually; it is pure magic. Everyone who is playing at that time knows where everybody's going. We all feel like one; wherever you go, everyone feels that's where we should go. I would know if Paul was going to do something, or if George was going to raise it up a bit, or John would double, or we'd bring it down. I usually play with my eyes closed, so you would know when things like that were happening . . . you've got to trust each other.
Huey Lewis (p. 179):
I find it more of a group experience for me. You look around and all of a sudden the song is playing and singing itself. It’s just like a wave that you ride. It’s tremendously exhilarating; it doesn’t take any energy and you look around and say, ‘Yep, this is it!’ It happens quite often but not for long periods of time. Almost at every gig that will happen somewhere for a fleeting moment. Some gigs it happens more often than others, and those are the good gigs.
Eric Clapton (p. 185):
It's a massive rush of adrenaline which comes at a certain point. Usually it's a sharing experience; it's not something I could experience on my own. . . . other musicians . . . an audience . . . Everyone in that building or place seems to unify at one point. It's not necessarily me that's doing it, it may be another musician. But it's when you get that completely harmonic experience, where everyone is hearing exactly the same thing without any interpretation whatsoever or any kind of angle. They're all transported toward the same place. That's not very common, but it always seems to happen at least once a show.
I don’t for a moment think these musicians are all describing the same thing. But just how many different things they’re describing, that I don’t know. Branford Marsalis says, “I feel that musicians who say it happens every time they play are full of shit.” Eric Clapton says that “it always seems to happen at least once a show.” Are they talking about the same thing? How could you tell?

Whatever it is, this family of musical experiences, it visits us in this devastating performance of Leonard Bernstein’s “One hand, one heart” by Jose Carreras and Kiri Te Kanawa.

Wednesday, November 24, 2021

Flowers and dogs [Japan]

Tuesday, November 16, 2021

Evening: Hoboken, Manhattan

Neural architecture search (NAS): A ML technique for automating the creation of neural networks

Wednesday, November 10, 2021

Understanding the role of individual units in a deep neural network

Abstract from the linked article:

Deep neural networks excel at finding hierarchical representations that solve complex tasks over large datasets. How can we humans understand these learned representations? In this work, we present network dissection, an analytic framework to systematically identify the semantics of individual hidden units within image classification and image generation networks. First, we analyze a convolutional neural network (CNN) trained on scene classification and discover units that match a diverse set of object concepts. We find evidence that the network has learned many object classes that play crucial roles in classifying scene classes. Second, we use a similar analytic method to analyze a generative adversarial network (GAN) model trained to generate scenes. By analyzing changes made when small sets of units are activated or deactivated, we find that objects can be added and removed from the output scenes while adapting to the context. Finally, we apply our analytic framework to understanding adversarial attacks and to semantic image editing.

Thursday, November 4, 2021

On the problem of adversarial attacks in machine learning

Wednesday, November 3, 2021

Facebook's Metaverse, Ho Hum, WTF!

From Ethan Zuckerman, Hey, Facebook, I Made a Metaverse 27 Years Ago, The Atlantic, October 29, 2021:

The metaverse Zuckerberg shows off in his video doesn’t have to solve those problems. He’s promising future technologies that are five to 10 years off. But it still looks like junk. The fire in his fireplace is a roughly rendered glow. His superhero secret lair looks out over a paradise island that’s almost entirely static. There’s the nominal motion of waves, but none of the foliage moves. It’s tropical wallpaper pasted to virtual windows. The sun is setting behind Zuckerberg’s left shoulder, but he’s being lit from the right front. Even with a bajillion dollars to invest in a video to relaunch and rename his company, Zuckerberg’s team is showing just how difficult it is to create a visually believable virtual world.

But that’s not the problem with Zuckerberg’s metaverse. The problem is that it’s boring. The futures it imagines have been imagined a thousand times before, and usually better. Two old men chat over a chessboard, one in Barcelona, one in New York, much as they did on Minitel in the 1980s. There’s virtual Ping-Pong and surfing, you know, like on a Wii. You can watch David Attenborough nature documentaries, like you do on Netflix. You can videoconference with your workmates … you know, like you do every single day.

Zuckerberg isn’t building the metaverse because he has a remarkable new vision of how things could be. There’s not an original thought in his video, including the business model. Thirty-eight minutes in, Zuckerberg gets serious, talking about how humbling the past few years have been for him and his business. Remember, he’s not humbled by the problem of Russian disinformation, or the spread of anti-vax misinformation, or the challenge of how Instagram affects teen body image. No, he’s humbled by how hard it is to fight against Apple and Google. [...]

Facebook can claim originality in at least one thing. Its combination of scale and irresponsibility has unleashed a set of diverse and fascinating sociopolitical challenges that it will take lawmakers, scholars, and activists at least a generation to fix. If Facebook has learned anything from 17 years of avoiding mediating those conflicts, it’s not apparent from the vision for the metaverse, where the power of human connection is celebrated as uncritically as it was before Macedonian fake-news brokers worked to sway the 2016 election.

How will a company that can block only 6 percent of Arabic-language hate content deal with dangerous speech when it’s worn on an avatar’s T-shirt or revealed at the end of a virtual fireworks display? Add monetization to the problems of content moderation—who gets to make money off a digital hate-speech T-shirt?—and Facebook’s oversight board is going to be very, very busy. [...]

The metaverse isn’t about building perfect virtual escape hatches—it’s about holding a mirror to our own broken, shared world. Facebook’s promised metaverse is about distracting us from the world it’s helped break.

Tuesday, November 2, 2021

Microsoft's Universal Image Language Representation model