Oct. 25th, 2023

Some stream-of-consciousness thoughts about consciousness in AI

A little late, but I noticed a lot of people been talking more about consciousness in AI since language models got better, like that guy from Google at some point and other people. Also, remembering a lot of discussions after Neurips last year when David Chalmers gave a keynote about what are the necessary conditions for consciousness and possible ways to test for some of those, and the others we don't really have an ability to test at this time.

Regardless of the "utility" of conscious AI, I think it's really interesting. Probably one of the most interesting things I could think of is exploring what makes our consciousness actually work and how to make one (other than having babies I suppose...different kind of interesting). However, it is one topic that really viscerally scares me. And no, not in the existential risk type of way that some people are focused on, for some good and some iffy reasons. I don't think most of those reasons particularly require consciousness anyway.

What I'm afraid of: Actually understanding the building blocks of consciousness feels like it would change my personal experience of life somehow. Like if we knew something about how conscious organisms experience time, and that we're able to change that. Or, that it's possible to construct situations where all outside interactions with the conscious entity seem like it's "one" entity but in fact is made of several conscious entities (maybe some having a worse time than others!). What if we just have an additional silent observer consciousness with us sharing our brains who "has no mouth and must scream". Or beyond this, many people's assumptions of an entirely material world might be disproven. We have a situation where we know if you change a brain, you can change the conscious experience of someone. But no matter how much physical info you have about a person's brain, you don't know what they're actually experiencing, because consciousness is inherently, definitionally subjective, and that's the hard problem of consciousness. I am of the view that the hard problem of consciousness is indeed a problem, that philosophical zombies are conceivable, etc. and haven't seen an argument yet to convince me otherwise.

So then we have a causal connection from the "real (material) universe" to what goes on in somebody's conscious experience. We have the "real world", and we got the whole "consciousness world" where people's qualia live I guess. That world: what rules govern it? Does it have rules? What can change? Different people have different experiences -- is that just because of different functional components? Many questions along these lines and probably more relevant ones undoubtedly have been asked by people a lot more deep into {the philosophical literature, spirituality, theology, neuroscience, etc.} than me.

But anyway, does the connection go in one direction? Are we so sure? And are the connections one-to-one? Is it possible to change someone's consciousness by somehow effecting a change in the "consciousness world" without necessarily affecting anything in the material world in the normal way to cause a change in consciousness? Probably not I guess but...why?

Anyway thanks for reading my blog :)

Protein ML Research

Thoughts about proteins as of Nov. 22, 2022

Houston, we got prions

Previous thoughts about proteins

Thoughts about proteins as of Dec. 8th, 2021

I'm working on a project to predict protein function descriptions in natural language, and focusing on evaluating the functional descriptions in an automated way. I'm trying to choose a good metric for this, starting my search with BLEU and other measures used in machine translation, and measures mentioned in this paper. If you know a lot about such measures, contact me! For this problem, compared to machine translation, there are differences in the assumptions of what constitutes a good match for a pair of descriptions, and how do we score a set of descriptions with a set of functions for a particular protein.

Prior thoughts

I've been working on function/fold/class discovery for proteins recently. I'm thinking about neural network-based clustering algorithms, though I know there are possibly much better ways to approach class discovery for proteins (probabilistic programming, energy based models). I want to learn more about those better approaches, but I still think it's worth exploring adapting the new techniques developed for unsupervised image classification for proteins just to see how they'd do.

Work surrounding mine

Some related work that I think is interesting.

• Vilnis, Luke, and Andrew McCallum. "Word representations via gaussian embedding." arXiv preprint arXiv:1412.6623 (2014).
• Mikolov, Tomas, et al. "Efficient estimation of word representations in vector space." arXiv preprint arXiv:1301.3781 (2013).
• Grover, Aditya, and Jure Leskovec. "node2vec: Scalable feature learning for networks." Proceedings of the 22nd ACM SIGKDD international conference on Knowledge discovery and data mining. 2016.
• Radford, Alec, et al. "Learning transferable visual models from natural language supervision." arXiv preprint arXiv:2103.00020 (2021)
• Van Gansbeke, Wouter, et al. "Learning to classify images without labels." arXiv preprint arXiv:2005.12320 (2020).
• Caron, Mathilde, et al. "Deep clustering for unsupervised learning of visual features." Proceedings of the European Conference on Computer Vision (ECCV). 2018.
• Singh, Rohit, Jinbo Xu, and Bonnie Berger. "Global alignment of multiple protein interaction networks with application to functional orthology detection." Proceedings of the National Academy of Sciences Sep 2008, 105 (35) 12763-12768; DOI: 10.1073/pnas.0806627105
• Ashburner, Michael, et al. "Gene ontology: tool for the unification of biology." Nature genetics 25.1 (2000): 25-29.
• Zhou, N., Jiang, Y., Bergquist, T.R. et al. The CAFA challenge reports improved protein function prediction and new functional annotations for hundreds of genes through experimental screens. Genome Biol 20, 244 (2019). https://doi.org/10.1186/s13059-019-1835-8

IsoRank (Singh et al. 2008)

IsoRank is a global network alignment algorithm that can be used to detect functionally similar proteins between two interaction networks. It involves two main steps:

1. Solving an eigenvalue equation to compute a functional similarity score matrix R between all pairs of proteins between the two networks
2. Extracting a set of high-scoring and mutually consistent matches from the R matrix.

SCAN (Semantic clustering by Adopting Nearest neighbors) (Van Gansbeke et al. 2020)

SCAN is a neural network-based clustering algorithm that has been used to classify images in an unsupervised way. This one involves three main steps:

1. Self-supervised learning of image features using a neural network, giving a k-nearest neighbors graph in this learned feature space
2. Train a softmax layer on top of the embeddings of the previous model using a semantic clustering loss function which enforces neighboring samples in the k-nearest neighbors graph to be in the same class
3. Additional training using pseudo-labels extracted from the model with strongly-augmented images in order to refine and increase confidence in predictions