Philip Castiglione x Cusack Lab

For over a decade, the two topics that have fascinated me most are the human mind and artificial intelligence.

The science-fiction potential of "AGI" and "super-intelligence" can easily capture the imagination. Software engineers in particular seem prone to such fantasy, often discounting the spectacularly complex mysteries of human cognition.

Despite recent breakthroughs in transformer-based and generative AI models, claims that we are approaching AGI ("scale is all you need") remain unpersuasive.3🔗 Situational Awareness — Leopold Aschenbrenner

With organisations like DeepMind leading the way, it seemed for a while like industry was committed to basic research and scientific progress. Further, backed by huge investment, loose reins and no bureaucracy.

However, private labs have recently disappointed:

1. Hyperfocus of intellectual effort and resources dedicated to limited lines of inquiry
2. The boom to bust investment / hype cycle of capital
3. An observable transition towards narrow commercial applications and extracting return on investment

As a result, I have turned towards the more rigorous and long-term oriented sphere of academic research.

Discovering brilliant people developing knowledge at the intersection of cognitive science, artificial intelligence and neuroscience has been a revelation.

I am now seeking a PhD in Cognitive Computational Neuroscience.

My ultimate goal is to contribute to our understanding of cognition so that tools with properties of intelligence can be employed for the benefit of humanity.

More specifically, the focus of my academic interest is on using techniques and lessons from neuroscience and artificial intelligence to explore cognition, which in turn will benefit our understanding of the brain and the sophistication of our models of intelligence. Foundational improvements to AI models like deep neural networks can be translated to technological applications which push forward the frontier of science.4🔗 Google DeepMind — Alphafold, 5 🔗 arXiv:2306.01495 , 6 🔗 World Modellers

I expect my research agenda will develop over time as I explore the existing research landscape and begin to make contributions. Pragmatic considerations including collaboration, publishing and funding opportunities will naturally have some influence.

I am also optimistic that there will be some room for big picture thinking and novel insights along the way.

So far, I have been immersing myself in theory and learning to read research papers (see What I am doing now below).

The starting point in my research career will be to gain experience in research and to build deep expertise in cognitive science, neuroscience, and the AI/ML techniques useful for their study.

I expect a strong starting point will be using my existing engineering skills to build AI models and pipelines while I build domain knowledge in other areas.

A PhD Studentship at the Cusack Lab working on the FreezeMotion project would be an ideal way to kick off my research career.

Continuing to learn neuroscience:

• Reading Netter's Atlas of Neuroscience - it's beautiful
• Various online educational videos

Watching the videos from CCN conference youtube channel (Clíona put me onto those!)

Teaching myself the pytorch machine learning framework popular in research

Extending on some topics of interest:

• Cognitive Science (A/Prof Fred Cummins, DCU)
• Deep Learning theory (Prof Andrew Ng, Stanford)
• Machine Learning foundations - math (Prof Andrew Ng, Stanford)

Practicing reading and summarising papers:

• Attention Is All You Need
• The remarkable, yet not extraordinary, human brain as a scaled-up primate brain and its associated cost
• The Platonic Representation Hypothesis
• Principles of Philosophy - René Descartes
• Contrastive Learning of Meaningful Object Associations from Temporal Co-occurrence Patterns in Naturalistic Movies
• Helpless infants are learning a foundation model
• Convolutional architectures are cortex-aligned de novo
• The Philosophy And Science Of Predictive Processing
• Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet