IQ-NET: A Holistic and Rapid Framework for Profiling the Intrinsic Aptitude of Neural Network Architectures
Yasser Sajjadi · 2025
Traditional neural network evaluation, rooted in resource-heavy benchmarks like GLUE [2] or ImageNet [1], is like reading only the final chapter of a book-it reveals outcomes but misses the story of a model's intrinsic strengths and weaknesses. IQ-NET revolutionizes this paradigm by profiling architectures in minutes using lightweight, synthetic probe tasks across text, image, audio, and video domains [5, 12]. Assessing 19 theoretically grounded metrics-from efficiency and scalability to reasoning and robustness-IQ-NET unveils a vivid "personality profile" for each architecture [6]. We evaluated five models (Zarvan, Transformer, LSTM, GRU, CNN), revealing striking contrasts: the attention-based Zarvan demonstrates exceptional multi-domain versatility, leading with a commanding final IQ score of 93.05, while recurrent models like GRU dominate stateful, algorithmic reasoning. IQ-NET's rapid, CPUbased approach empowers researchers to select task-optimal models, pinpoint improvement areas, and spark architectural innovation. All data and code are available at https://github.com/systbs/iq-net [4].