spotlight poster @ ICLR!

Date:

Fede’s work got selected as a spotlight poster @ the International Conference on Learning Representations!

Adolfi F, Vilas MG and Wareham T (2025). The Computational Complexity of Circuit Discovery for Inner Interpretability.

Available here.

Many proposed applications of neural networks in machine learning, cogni- tive/brain science, and society hinge on the feasibility of inner interpretability via circuit discovery. This calls for empirical and theoretical explorations of viable al- gorithmic options. Despite advances in the design and testing of heuristics, there are concerns about their scalability and faithfulness at a time when we lack under- standing of the complexity properties of the problems they are deployed to solve. To address this, we study circuit discovery with classical and parameterized com- putational complexity theory: (1) we describe a conceptual scaffolding to reason about circuit finding queries in terms of affordances for description, explanation, prediction and control; (2) we formalize a comprehensive set of queries for mech- anistic explanation, and propose a formal framework for their analysis; (3) we use it to settle the complexity of many query variants and relaxations of practical in- terest on multi-layer perceptrons. Our findings reveal a challenging complexity landscape. Many queries are intractable, remain fixed-parameter intractable rela- tive to model/circuit features, and inapproximable under additive, multiplicative, and probabilistic approximation schemes. To navigate this landscape, we prove there exist transformations to tackle some of these hard problems with better- understood heuristics, and prove the tractability or fixed-parameter tractability of more modest queries which retain useful affordances. This framework allows us to understand the scope and limits of interpretability queries, explore viable options, and compare their resource demands on existing and future architectures.

Recommended citation: Adolfi F, Vilas MG and Wareham T (2025). The Computational Complexity of Circuit Discovery for Inner Interpretability.

You May Also Enjoy

New preprint!

Date:

New preprint led by Jiajie!

{{ page.citation }}

Available here.

Spoken language is a rapid sequence of auditory information that can incorporate complicated internal structures. Prediction and chunking are two possible mechanisms for the brain to process speech efficiently, and they are often viewed as separate or even opposing mechanisms. Here we investigate whether the two mechanisms interact and hypothesize that the chunk (constituent) structure in spoken language modulates how the brain predicts basic linguistic items, e.g., morphemes, the basic unit of meaning in language. In three magnetoencephalography (MEG) experiments, we characterize the neurally manifested prediction of morphemes using the neural response to morpheme surprisal, and we analyze how this response is modulated by chunks, i.e., the linguistic constituent structure. We demonstrate that the MEG surprisal response is significantly stronger for morphemes within an ongoing chunk than morphemes across a chunk boundary. This chunk-boundary effect on morpheme prediction is further modulated by the certainty of a chunk boundary. The findings are then confirmed by analyzing a dataset of ECoG responses to English narratives. These results establish that the brain employs a chunk-based prediction mechanism and more precisely predicts sequential items within a chunk.

New paper!

Date:

New paper led by Yue!

{{ page.citation }}

Available here.

Ample evidence demonstrates synchronized neurophysiological activity to the temporal regularities of sounds. This phenomenon has been proposed to reflect neural responses to onset edges in acoustic signals. Here, we examine listeners’ ability to behaviorally synchronize to stimuli with sharp or gradual onsets. In two experiments, we show that while the sharp amplitude onsets facilitate temporal phase alignment between participants’ behavioral output and the stimulus envelope, sharp onsets are not essential for tracking the rate of auditory rhythms in the acoustic input. The dissociation between phase and rate tracking suggests distinct underlying neural mechanisms that are separately modulated.

lab @ Cognitive Neuroscience Society annual meeting

Date:

Some members of the lab are at CNS! Come to following posters to chat. 🙂 

  • Poster C137: Phoebe Chen – Understanding spontaneous false memory in the naturalistic recall of narratives
  • Poster E102: Jiajie Zou – Chunking Constrains Prediction during Language Comprehension
  • Poster E104: Leonardo Zeine – Differentiating endogenous and entrained delta-band rhythms in the brain: a naturalistic story-listening experiment
  • Poster F61: Yue Sun – Complex impact of stimulus envelope on motor synchronization to sound
  • Poster F63: Berfin Bastug – Sensorimotor Engagement Facilitates Regularity Detection During Auditory Scene Analysis

New paper!

Date:

New paper led by Matthias!

Grabenhorst M, Poeppel D, Michalareas G (2025). Neural signatures of temporal anticipation in human cortex represent event probability density. Nature Communications.

Available here.

Temporal prediction is a fundamental function of neural systems. Recent results show that humans anticipate future events by calculating probability density functions, rather than hazard rates. However, direct neural evidence for this hypothesized mechanism is lacking. We recorded neural activity using magnetoencephalography as participants anticipated auditory and visual events distributed in time. We show that temporal anticipation, measured as reaction times, approximates the event probability density function, but not hazard rate. Temporal anticipation manifests as spatiotemporally patterned activity in three anatomically and functionally distinct parieto-temporal and sensorimotor cortical areas. Each of these areas revealed a marked neural signature of anticipation: Prior to sensory cues, activity in a specific frequency range of neural oscillations, spanning alpha and beta ranges, encodes the event probability density function. These neural signals predicted reaction times to imminent sensory cues. These results demonstrate that supra-modal representations of probability density across cortex underlie the anticipation of future events.