Similarity-Based Reasoning, Raven's Matrices, and General Intelligence

Authors: Can Serif Mekik, Ron Sun, David Yun Dai

IJCAI 2018 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental This paper presents a model tackling a variant of the Raven's Matrices family of human intelligence tests along with computational experiments.
Researcher Affiliation Academia Can Serif Mekika , Ron Suna, David Yun Daib a Rensselaer Polytechnic Institute b State University of New York at Albany
Pseudocode Yes 3.3 Algorithm To put everything together, in response to a matrix with 8 alternative answers (e.g., Figure 1a) the model proceeds as follows: 1. Using a convolutional neural network, subjective probabilities are obtained for each individual feature in each matrix sequence and each alternative sequence. 2. Feature distributions for matrix sequences are combined, row-wise and column-wise, using geometric means (Equations 4 and 5). 3. For each alternative, the similarity of the 2 resulting alternative sequences and the 2 combined matrix sequences is computed, with regard to each feature, row-wise and column-wise, through relative entropy (Equation 7). 4. The 8 resulting similarity measures are used to stochastically select an answer out of the 8 alternatives, using the Boltzmann distribution (Equation 3).
Open Source Code Yes 1 Experimental notes, data and code are available at https://osf.io/7fy2d/.
Open Datasets Yes The model was tested on a set of matrix problems generated using the Sandia Matrix Generation Tool [Matzen et al., 2010].
Dataset Splits No The paper states 'A total of 1480 input-output pairs (20 from each of 74 matrices) were used to train the model' and mentions 'test items', but does not explicitly provide details about a separate validation split or its size/percentage. No cross-validation setup is mentioned for splitting data.
Hardware Specification No The paper describes the neural network architecture and training procedure but does not provide any specific details about the hardware (e.g., CPU, GPU models, memory) used for running the experiments.
Software Dependencies No The paper mentions algorithms like ADADELTA but does not list specific software dependencies with version numbers (e.g., Python, TensorFlow, PyTorch versions).
Experiment Setup Yes The network was trained using the ADADELTA algorithm [Zeiler, 2012] for 24000 epochs with learning parameters 𝜌 = .6 and 𝜖 = 10i)Å. The Xavier method [Glorot and Bengio, 2010] was used for network weight initialization. Training samples were shuffled before being split into 20 mini-batches of 74 input-output pairs. Mini-batch order was shuffled at the start of each epoch, and, every 500 epochs, network weights were recorded. A cross-entropy error measure augmented with 𝐿) (𝜆) = .003) and 𝐿* (𝜆* = .003) weight penalties was used to compute weight updates.