At that earlier time, datasets were so small that hand-engineered methods worked better than convolutional nets.

'That earlier time' means the period before large datasets, roughly the 1990s to the late 2000s, before ImageNet-scale benchmarks. Judge as of 2021 against the historical record. True: benchmark results and historical accounts from that period show that on the main tasks (especially vision benchmarks), hand-engineered feature methods generally performed as well as or better than convolutional nets, and that small dataset size is commonly given as a main reason. Mostly true: hand-engineered methods generally won, but the evidence attributes this only partly to dataset size, or convnets did win on some small-data tasks such as digit recognition. Mixed: the two approaches were roughly level, or which one won depended heavily on the task. Mostly false: convolutional nets were already competitive or better on most benchmarks, and hand-engineered methods led only in narrow cases. False: convolutional nets generally outperformed hand-engineered methods in that period.

Set before the research began.

1. E1 Limitations of Deep Learning for Vision, and How We Might Fix Them The Gradient · 9 February 2019

   “While most of the basic ideas of deep learning were already developed during the second wave, their power could not be unleashed until large datasets and powerful computers (GPUs) became available.”

2. E2 Kernel Descriptors for Visual Recognition University of Washington (Bo, Ren, Fox; NIPS paper)

   “Both our gradient kernel descriptor and shape kernel descriptor are superior to CDBN by a large margin.”

3. E3 Deep learning Nature (LeCun, Bengio & Hinton review, 2015) · 28 May 2015 Own record

   “By the late 1990s this system was reading over 10% of all the cheques in the United States.”

E1 links deep learning's progress in general to large datasets and GPUs, and E2 reports one undated result in which kernel descriptors beat a convolutional deep belief network. E4 does not show that convnets succeeded on small-data tasks. None of the evidence shows whether hand-engineered methods generally outperformed convolutional nets on the main benchmarks of the 1990s to late 2000s, so the criteria cannot be settled.