Medical applications are critical because they directly impact human lives. These applications require precise, controlled results akin to those provided by human specialists.
ML
“In machine learning, models were designed either from first principles or based on reasonable heuristics. As a result, users typically had a clear understanding of the features used and the decision-making process.”
These models exhibit an explicit decision boundary. In linear models (such as logistic regression and support vector machines), the decision boundary corresponds to a hyperplane defined by the model parameters. For K-Nearest Neighbor (KNN), the decision boundary(s) form a manifold equidistant from different clusters. In the case of decision trees (DT), the decision boundary(s) follow the manifold specified by the branching structure. Secondly, these models are particularly effective when working with small datasets.
DL
DL’s to obtain models with human-level performance in specialized tasks. Compared to classical Machine Learning, deep learning models are often considered ‘black boxes.’ In these models, learned features are implicit and challenging to interpret. Additionally, the decision boundaries are not explicit, making it nearly impossible to measure the distance between observed data points and these boundaries.
Empirical evidence suggests that deep learning models are often ill-calibrated, meaning that their initial estimated confidence levels do not always align with their true confidence.
LLM
In the era of large language models, excitement abounds. Training these models typically involves three stages: pre-training, fine-tuning, and in-context learning. During pre-training, self-supervised learning occurs, either in a causal or non-causal manner, allowing the training process to leverage abundant unlabeled data.
The first characteristic of Large Language Models (LLMs) is their extensive datasets. Through self-supervised learning, LLMs are exposed to significantly more data points than previous models, which allows for a more generalized in-distribution understanding. However, since these data points are unlabeled, the concept of in-distribution differs from the Deep Learning (DL) era, where both data and labels are available. This raises questions about how unlabeled data can best be utilized to improve performance on downstream tasks. An intriguing concept related to this is parametric memory, which views LLMs as large “memorizers” where knowledge consists of observed data points.
The second characteristic is model size. LLMs are much larger than earlier models, making the use of computationally intensive techniques impractical.
The third characteristic is task generalization. Unlike models specialized for single tasks, LLMs, often associated with foundation models, are expected to learn general features from vast amounts of unlabeled data. These general or semantic features are beneficial for various related downstream tasks.
In machine learning (ML), deep learning (DL), and large language models (LLMs), several inherent issues significantly affect their reliability, especially in critical applications like medical science. While they hold great promise for advancing medical science, their issues of bias, opacity, and hallucination pose significant challenges. Addressing these challenges requires a multifaceted approach involving better data practices, algorithmic transparency, and rigorous validation to ensure that these models can be safely and effectively integrated into medical critical applications.
Therefore, at Curedis.ai we use best of all worlds, always mindful of the impact our interpretations and decisions have on the health and lives of the recipients.