The machinery of attention: a reading list from IISc’s Centre for Neuroscience

One idea runs through most of this work. A behavioural response, deciding whether a target was there, can be split into two independent quantities: perceptual sensitivity, how well you actually saw it, and choice bias, how willing you were to say yes. The lab uses signal detection theory to pull these apart, then asks which one attention, expectation or reward is really moving, and where in the brain each lives. The answers turn out to be surprisingly separable.

Two components of attention: sensitivity and bias

• Distinct neural bases of subcomponents of the attentional blink

  eLife’24

  When two targets arrive in quick succession the second is often missed, the attentional blink. Using a multialternative task and a signal detection model, the blink is shown to cut perceptual sensitivity while leaving choice bias intact, with parieto-occipital N2p and P3 amplitudes tracking the detection deficit and long-range high-beta (20–30 Hz) frontoparietal coherence tracking the discrimination deficit.

• Dissociable components of attention exhibit distinct neuronal signatures in primate visual cortex

  Science Advances’24

  In monkeys performing a task that separates where the eyes plan to move from where covert attention sits, the two components come apart cleanly. Bias rose at the focus of covert attention and tracked V4 firing rates, while sensitivity rose at the planned saccade target and tracked reduced correlated variability between pairs of neurons.

• The right posterior parietal cortex mediates spatial reorienting of attentional choice bias

  Nature Communications’24

  To test whether the dorsal right posterior parietal cortex (rPPC) causally drives attention, participants did a cued task under either inhibitory repetitive TMS or 40 Hz transcranial alternating current stimulation of the rPPC. Both selectively altered the bias component, not sensitivity, reducing bias toward the uncued location, pinning spatial reorienting of bias on the rPPC.

Telling attention, expectation and reward apart

• Attention, not expectation, governs spatial selection by probabilistic cueing

  NeuroImage’25

  A probabilistic cue could bias where you look through attention or through expectation; this teases them apart with signal detection models and concurrent EEG. Posner and relevance (attention) cues, but not probability (expectation) cues, modulated SSVEP amplitude and alpha-band power, and representational similarity analysis showed Posner cueing looked like attention, so the effect is attention rather than expectation.

• Reward expectation yields distinct effects on sensory processing and decision making in the human brain

  PLOS Biology’25

  Does expecting a reward sharpen perception, or just tilt the decision? Reward expected across spatial locations raised sensitivity and engaged the usual neural attention markers, whereas reward tied to a specific choice moved the decision criterion (bias) without those markers, so reward shapes seeing and deciding through separable mechanisms.

Deep learning for medical images

• SASHA: Sequential attention based sampling for histopathological analysis

  NeurIPS’25

  Whole-slide pathology images are gigapixel, labelled only at the slide level, with the tell-tale regions sparse. SASHA learns features with a lightweight hierarchical attention multiple-instance model, then zooms into just 10–20% of the high-resolution patches, matching full-resolution accuracy at a fraction of the compute and beating other sparse-sampling methods.

• CRAFT: Semi-supervised deep transfer for regression without domain alignment

  ICCV’25

  Adapting a regression model to a new domain is hard when the source data cannot be kept, for privacy or storage, and labels are scarce in the new one. CRAFT regularizes with a contradistinguisher instead of aligning representations, improving RMSE by up to 9% over fine-tuning and beating four source-free methods by more than 3% on EEG gaze and MRI brain-age prediction.

The recurring move is to refuse to treat “attention” as one thing. Split behaviour into sensitivity and bias first, and the neuroscience gets sharper: different scalp signatures, different cortical areas, even different effects of reward and expectation. The two computational papers carry the same instinct into imaging, where a learned attention decides which slivers of a slide are worth the compute. Full details are in each linked paper, and the lab’s wider output is on its site.