1According to Marr's levels of analysis, what is the focus of the 'computational problem' level?
This level focuses on the 'why' of a behavior, typically framed as an optimization problem (e.g., Bayesian models).
2According to Marr's levels of analysis, what is the focus of the 'algorithmic problem' level?
This level focuses on the 'which steps' are achieved to produce a behavior, typically addressed by process models.
3What is the focus of the 'implementation problem' in Marr's levels of analysis?
This level focuses on how a process is implemented in the brain, typically addressed by neural models like machine learning.
4Newell's time scales of action are NOT based on the total duration of a task, but rather on the level at which the task is _____.
modelled
5In Newell's time scales of action, what does the 'Biological Band' typically correspond to in terms of time?
Milliseconds (ms).
6In Newell's time scales of action, what does the 'Cognitive Band' typically correspond to in terms of time?
Seconds (s).
7What time scale does the 'Rational Band' in Newell's framework typically represent?
Minutes (min).
8What time scale does the 'Social Band' in Newell's framework typically represent?
Days, weeks, or months.
9In the classification of cognitive models, what is the key characteristic of a 'top-down' approach?
It is theory-informed, moving from theory to experiment.
10In the classification of cognitive models, what is the key characteristic of a 'bottom-up' approach?
It is data-driven, moving from experiment towards theory.
11What is the distinction between a theory-driven ('white box') model and a data-driven ('black box') model?
A theory-driven model aims to explicitly model the underlying process, while a data-driven model focuses on mapping inputs to outputs without necessarily modeling the intermediate steps.
12What is the primary difference between a 'stand-alone' model and a 'cumulative' model?
A stand-alone model is built for a single purpose, while a cumulative model builds on preceding theory.
13In cognitive modeling, what does it mean for a model to be 'embodied'?
It has its own sensory and action system to interact with the environment.
14In cognitive modeling, what does it mean for a model to be 'situated'?
It acquires all information about the environment from its own perspective through its sensory systems.
15What is the proposed interdisciplinary research domain for studying collaboration dynamics in sociotechnical systems, as advocated by Gupta et al.?
Collective Human-Machine Intelligence (COHUMAIN).
16In the context of COHUMAIN, a _____ specifies the underlying infrastructure, components, and functional processes for a multi-agent, sociotechnical system.
sociocognitive architecture
17What is the first of the four core problems (P1) for sociocognitive architectures in COHUMAIN?
How individual members perceive and represent each other's mental states (e.g., goals, beliefs) to support the emergence of collective cognition.
18What is the second core problem (P2) for sociocognitive architectures in COHUMAIN?
How members engage in trustworthy interactions to dynamically align their diverse mental states and select joint priorities.
19What is the third core problem (P3) for sociocognitive architectures in COHUMAIN?
How individual members perceive and represent each other's cognitive resources (e.g., knowledge, skills) to facilitate collective cognition.
20What is the fourth core problem (P4) for sociocognitive architectures in COHUMAIN?
How members develop shared norms of interaction to dynamically coordinate interdependent actions and ensure efficient resource use.
21What are the two necessary features that any sociocognitive architecture for COHUMAIN must specify?
Mechanisms for (1) between-member metacognition and (2) the development of engagement rules for coordinated action and trust.
22What is the name of the sociocognitive architecture proposed by Gupta & Woolley that models collective intelligence as three interlocking systems?
The Transactive Systems Model of Collective Intelligence (TSM-CI).
23What are the three coregulatory, sociocognitive systems that form the TSM-CI?
Transactive Memory System (TMS), Transactive Attention System (TAS), and Transactive Reasoning System (TRS).
24What is the primary function of a Transactive Memory System (TMS)?
To coordinate distributed knowledge and skills via meta-memory processes (beliefs about who knows what).
25How can an AI agent augment a team's Transactive Memory System (TMS)?
By identifying misused or underutilized experts to optimize team composition and task allocation.
26What is the primary function of a Transactive Attention System (TAS)?
To coordinate distributed attentional focus and priorities via meta-attention processes.
27How might an AI agent enhance a group's Transactive Attention System (TAS)?
By identifying workload conflicts and bottlenecks to help coordinate handoffs and manage collective attention.
28What is the primary function of a Transactive Reasoning System (TRS)?
To coordinate and align diverse goals and preferences via meta-reasoning processes.
29What is one way an AI agent can support a team's Transactive Reasoning System (TRS)?
By nudging information in the choice environment to facilitate the human members' joint decision-making.
30What cognitive architecture is presented as being particularly compatible with the TSM-CI for developing AI agents?
Instance-Based Learning Theory (IBLT).
31In Instance-Based Learning Theory (IBLT), how are decisions represented in memory?
As instances, which are triplets involving the state, actions, and utilities (S, A, U).
32What is the general term for AI agents that can predict the cognitive states (e.g., desires, intentions, beliefs) of other agents?
Theory of Mind (ToM) agents.
33What is HToMM an acronym for in the context of human-AI interaction?
Human Theory of Machine Mind (a human understanding a machine's cognition).
34What is MToHM an acronym for in the context of human-AI interaction?
Machine Theory of Human Mind (a machine modeling and predicting human cognitive states).
35At what level of abstraction are the 'neural network models' discussed by Storrs & Kriegeskorte typically situated?
At an intermediate level, between detailed biological simulations and highly abstract cognitive models.
36What two operations does a single unit in a typical neural network model perform on its inputs?
It computes a weighted sum of its inputs and passes the result through a nonlinear activation function.
37In a neural network, flexible knowledge is held by the _____, which are analogous to synaptic weights in the brain.
weights associated with connections between units
38What type of learning algorithm adjusts network weights to bring outputs closer to pre-specified desired outputs for a set of labeled input patterns?
Supervised learning.
39What is the name of the efficient algorithm used in supervised learning to compute the derivatives of the cost function with respect to each weight?
Backpropagation.
40What type of learning algorithm involves training a network on unlabeled data to model the statistics of its inputs?
Unsupervised learning.
41An _____ is a type of network that, through unsupervised learning, learns to compress its input into a lower-dimensional code and then reconstruct it.
autoencoder
42What type of learning algorithm involves a network outputting an action and learning to model the expected cumulative reward associated with possible actions?
Reinforcement learning.
43What is the term for a neural network architecture where units form a single processing hierarchy with no loops or feedback connections?
Feedforward architecture.
44What defines a 'recurrent' neural network architecture?
It contains one or more loops in its connection graph, such as top-down feedback, allowing its internal state to evolve over time.
45A feedforward network with only a single 'hidden' layer of units between its input and output is known as a _____ network.
shallow
46What term describes a neural network with more than one hidden layer intervening between its input and output units?
Deep network.
47What type of specialized network architecture, inspired by the mammalian visual cortex, is often used for visual tasks?
Deep convolutional neural networks (CNNs).
48What two key features of CNNs reduce the number of parameters and capture the prior belief that a feature can appear anywhere in an image?
Spatially restricted receptive fields and shared weights.
49What specialized unit type in recurrent networks can dynamically retain recent information for as long as it is useful in the current task?
Long short-term memory (LSTM) units.
50What is the purpose of Representational Similarity Analysis (RSA) when comparing a DNN model to the brain?
To use a stimulus-by-stimulus dissimilarity matrix to characterize the representational space and compare model layers to brain regions.
51The ability of _____ to predict IT cortical representation outperformed shallow computer vision models.
deep supervised models
52What is one novel interrogation technique, possible with DNNs but not brains, that uses gradient-descent to create stimuli that elicit specific network activation patterns?
Stimulus optimization (or creating optimal stimuli).
53What is a major limitation of standard DNNs regarding their robustness, as demonstrated by their susceptibility to minute, human-imperceptible image perturbations?
They are fragile and can be tricked into drastic misclassifications (adversarial examples).
54From a biological and ecological perspective, what is a major criticism of supervised training that requires millions of labeled stimuli?
It is an ecologically unrealistic requirement, as animal brains learn far more data-efficiently, often without explicit supervision.
55What influential neuroscience idea, where neural activity primarily encodes the difference between predicted and actual sensory input, offers a rich unsupervised training signal?
Predictive coding.
56In human-autonomy teaming (HAT), what is the term for the need for both humans' awareness of autonomous agents and vice versa for effective performance?
Shared situational awareness.
57What are the three factors of trustworthiness demonstrated in humans that are being examined for their applicability to human-AI trust?
Competence, benevolence, and integrity.
58According to Glikson and Woolley's review, research on human trust in AI has to date focused almost exclusively on dimensions related to _____, particularly reliability.
competence
59Work on explainable AI (XAI) is largely focused on promoting human-AI trust by reinforcing perceptions of an AI's ____.
competence
60Which basis of human-AI trust, related to perceptions that an agent is acting in the user's best interest, has received considerably less attention in research?
Benevolence and related affective bases of trust.
61What is the core problem that cognitive architectures and sociocognitive architectures aim to solve, respectively?
Cognitive architectures focus on how an autonomous agent perceives and acts, while sociocognitive architectures focus on how multiple agents collaborate and problem-solve together.